[Prim][PIR] support unbind op forward in prim pir

paddle/fluid/pir/dialect/op_generator/decomp_interface_gen_op_list.py

@@ -56,6 +56,7 @@
     "squeeze",
     "stack",
     "unsqueeze",
+    "unbind",
     "huber_loss",
 ]
 
@@ -96,6 +97,7 @@
     "squeeze",
     "stack",
     "unsqueeze",
+    "unbind",
     "huber_loss",
 ]
 

paddle/fluid/primitive/composite/composite.h

@@ -613,6 +613,23 @@ std::vector<Tensor> meshgrid_decomp(const std::vector<Tensor>& x) {
   return res;
 }
 
+template <typename T>
+std::vector<Tensor> unbind_decomp(const Tensor x, int axis) {
+  std::vector<Tensor> res;
+  if (axis < 0) {
+    axis = x.shape().size() + axis;
+  }
+  if (x.shape()[axis] == -1) {
+    PADDLE_THROW(phi::errors::Unimplemented("unbind axis must not be dynamic"));
+  }
+  size_t num = x.shape()[axis];
+  std::vector<Tensor> tmp = backend::split_with_num<T>(x, num, axis);
+  for (size_t i = 0; i < tmp.size(); i++) {
+    res.push_back(squeeze<T>(tmp[i], {axis}));
+  }
+  return res;
+}
+
 template <typename T>
 std::tuple<Tensor, Tensor, Tensor> layer_norm_decomp(
     const Tensor& x,

test/deprecated/legacy_test/test_unbind_op.py

@@ -167,6 +167,7 @@ def setAxis(self):
 
     def setUp(self):
         self._set_op_type()
+        self.prim_op_type = "comp"
         self.dtype = self.get_dtype()
         self.axis = 0
         self.num = 3
@@ -186,6 +187,7 @@ def setUp(self):
             'Out': [('out%d' % i, self.out[i]) for i in range(len(self.out))]
         }
         self.python_api = paddle.unbind
+        self.public_python_api = paddle.unbind
         self.python_out_sig = ['out%d' % i for i in range(len(self.out))]
 
     def get_dtype(self):
@@ -195,10 +197,12 @@ def _set_op_type(self):
         self.op_type = "unbind"
 
     def test_check_output(self):
-        self.check_output(check_pir=True)
+        self.check_output(check_pir=True, check_prim_pir=True)
 
     def test_check_grad(self):
-        self.check_grad(['X'], ['out0', 'out1', 'out2'], check_pir=True)
+        self.check_grad(
+            ['X'], ['out0', 'out1', 'out2'], check_pir=True, check_prim_pir=True
+        )
 
 
 class TestUnbindOp1(TestUnbindOp):
@@ -263,47 +267,73 @@ class TestUnbindOp1_Complex64(TestUnbindOp1):
     def get_dtype(self):
         return np.complex64
 
+    def test_check_output(self):
+        self.check_output(check_pir=True)
+
 
 class TestUnbindOp2_Complex64(TestUnbindOp2):
     def get_dtype(self):
         return np.complex64
 
+    def test_check_output(self):
+        self.check_output(check_pir=True)
+
 
 class TestUnbindOp3_Complex64(TestUnbindOp3):
     def get_dtype(self):
         return np.complex64
 
+    def test_check_output(self):
+        self.check_output(check_pir=True)
+
 
 class TestUnbindOp4_Complex64(TestUnbindOp4):
     def get_dtype(self):
         return np.complex64
 
+    def test_check_output(self):
+        self.check_output(check_pir=True)
+
 
 class TestUnbindOp1_Complex128(TestUnbindOp1):
     def get_dtype(self):
         return np.complex128
 
+    def test_check_output(self):
+        self.check_output(check_pir=True)
+
 
 class TestUnbindOp2_Complex128(TestUnbindOp2):
     def get_dtype(self):
         return np.complex128
 
+    def test_check_output(self):
+        self.check_output(check_pir=True)
+
 
 class TestUnbindOp3_Complex128(TestUnbindOp3):
     def get_dtype(self):
         return np.complex128
 
+    def test_check_output(self):
+        self.check_output(check_pir=True)
+
 
 class TestUnbindOp4_Complex128(TestUnbindOp4):
     def get_dtype(self):
         return np.complex128
 
+    def test_check_output(self):
+        self.check_output(check_pir=True)
+
 
 class TestUnbindFP16Op(OpTest):
     def setUp(self):
         paddle.disable_static()
         self.op_type = "unbind"
+        self.prim_op_type = "comp"
         self.python_api = paddle.unbind
+        self.public_python_api = paddle.unbind
         self.dtype = self.get_dtype()
         self.axis = 0
         self.num = 3
@@ -326,14 +356,16 @@ def get_dtype(self):
         return np.float16
 
     def test_check_output(self):
-        self.check_output(check_pir=True)
+        self.check_output(check_pir=True, check_prim_pir=True)
 
 
 class TestUnbindBF16Op(OpTest):
     def setUp(self):
         paddle.disable_static()
         self._set_op_type()
+        self.prim_op_type = "comp"
         self.python_api = paddle.unbind
+        self.public_python_api = paddle.unbind
         self.dtype = self.get_dtype()
         self.axis = 0
         self.num = 3
@@ -362,7 +394,7 @@ def _set_op_type(self):
         self.op_type = "unbind"
 
     def test_check_output(self):
-        self.check_output(check_pir=True)
+        self.check_output(check_pir=True, check_prim_pir=True)
 
     def test_check_grad(self):
         pass

test/ir/pir/cinn/symbolic/test_infer_sym_shape_unary_op.py

@@ -24,6 +24,7 @@
 
 import paddle
 import paddle.nn.functional as F
+from paddle.framework import core
 from paddle.static import InputSpec
 
 sys.path.append(dirname(dirname(__file__)))
@@ -807,6 +808,8 @@ def prepare_data(self):
         ]
 
     def test_eval_symbolic(self):
+        core._set_prim_forward_blacklist("pd_op.unbind")
+
         net = UnbindNet()
 
         for i in range(len(self.cases)):

test/prim/pir_prim/test_prim_sub_graph_dynamic_shape.py

@@ -156,6 +156,10 @@ def meshgrid_net(x, y):
     return paddle.meshgrid(x, y)
 
 
+def unbind_net(x):
+    return paddle.unbind(x)
+
+
 class TestPrimBase(unittest.TestCase):
     def setUp(self):
         np.random.seed(2023)
@@ -231,6 +235,19 @@ def setUp(self):
         self.tol = 1e-6
 
 
+class TestUnbind(TestPrimBase):
+    def setUp(self):
+        np.random.seed(2023)
+        self.dtype = "float32"
+        self.x_shape = [4, 5, 6]
+        self.init_x_shape = [4, 5, None]
+        self.x = np.random.random(self.x_shape).astype(self.dtype)
+        self.net = unbind_net
+        self.necessary_ops = "pd_op.unbind"
+        self.enable_cinn = False
+        self.tol = 1e-6
+
+
 class TestPrimFullLike(TestPrimBase):
     def setUp(self):
         np.random.seed(2023)